ROSEBUD: A Deep Fluvial Segmentation Dataset for Monocular Vision-Based River Navigation and Obstacle Avoidance.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Semantic image segmentation using neural networks is vital for autonomous navigation.
• Limited training data hinders application in complex water environments like rivers.
• Existing marine datasets are insufficient for fluvial navigation challenges.

Purpose of the Study:

• Introduce the River Obstacle Segmentation En-Route By USV Dataset (ROSEBUD).
• Provide a benchmark for surface navigation in complex fluvial scenes.
• Address the data scarcity for semantic segmentation in riverine environments.

Main Methods:

• Created and released the ROSEBUD dataset with 549 hand-annotated fluvial images.
• Evaluated state-of-the-art semantic networks trained on existing marine datasets.
• Assessed network generalization and performance on the ROSEBUD dataset.

Main Results:

• The ROSEBUD dataset presents a challenging baseline for fluvial obstacle detection.
• Networks trained on marine data show limitations in segmenting small obstacles in rivers.
• Further training on ROSEBUD significantly improves segmentation accuracy for fluvial scenes.

Conclusions:

• The ROSEBUD dataset is essential for advancing semantic segmentation in river navigation.
• Specialized training data is required for robust obstacle detection in complex inland waterways.
• This work facilitates improved autonomous navigation capabilities for unmanned surface vehicles in rivers.